Regression Trees Part1

Preparations

https://besjournals.onlinelibrary.wiley.com/doi/10.1111/j.1365-2656.2008.01390.x

Load the necessary libraries

library(gbm)         #for gradient boosted models
library(car)
library(dismo)
library(pdp)
library(ggfortify)
library(randomForest)
library(tidyverse)
library(gridExtra)
library(patchwork)

Scenario

Abalone are an important aquaculture shell fish that are farmed for both their meat and their shells. Abalone can live up to 50 years, although their longevity is known to be influenced by a range of environmental factors. Traditionally, abalone are aged by counting thier growth rings, however, this method is very laborious and expensive. Hence a study was conducted in which abalone growth ring counts were matched up with a range of other more easily measured physical characteristics (such as shell dimesions and weights) in order to see if any of these other parameters could be used as proxies for the number of growth rings (or age).

abalone

Format of abalone.csv data file

Read in the data

abalone = read_csv('../public/data/abalone.csv', trim_ws=TRUE)
glimpse(abalone)

## Rows: 4,177
## Columns: 10
## $ SEX          <chr> "M", "M", "F", "M", "I", "I", "F", "F", "M", "F", "F", "M…
## $ LENGTH       <dbl> 0.455, 0.350, 0.530, 0.440, 0.330, 0.425, 0.530, 0.545, 0…
## $ DIAMETER     <dbl> 0.365, 0.265, 0.420, 0.365, 0.255, 0.300, 0.415, 0.425, 0…
## $ HEIGHT       <dbl> 0.095, 0.090, 0.135, 0.125, 0.080, 0.095, 0.150, 0.125, 0…
## $ WHOLE_WEIGHT <dbl> 0.5140, 0.2255, 0.6770, 0.5160, 0.2050, 0.3515, 0.7775, 0…
## $ MEAT_WEIGHT  <dbl> 0.2245, 0.0995, 0.2565, 0.2155, 0.0895, 0.1410, 0.2370, 0…
## $ GUT_WEIGHT   <dbl> 0.1010, 0.0485, 0.1415, 0.1140, 0.0395, 0.0775, 0.1415, 0…
## $ SHELL_WEIGHT <dbl> 0.150, 0.070, 0.210, 0.155, 0.055, 0.120, 0.330, 0.260, 0…
## $ RINGS        <dbl> 15, 7, 9, 10, 7, 8, 20, 16, 9, 19, 14, 10, 11, 10, 10, 12…
## $ AGE          <dbl> 16.5, 8.5, 10.5, 11.5, 8.5, 9.5, 21.5, 17.5, 10.5, 20.5, …

abalone = abalone %>% mutate(SEX=factor(SEX))

Exploratory data analysis

ggplot(abalone) +
    geom_point(aes(y=AGE, x=RINGS))

## Very tight relationship

scatterplotMatrix(~RINGS + SEX +LENGTH+DIAMETER+HEIGHT+WHOLE_WEIGHT+MEAT_WEIGHT+GUT_WEIGHT+
                  SHELL_WEIGHT, data=abalone)

Fit the model

We will start by partitioning the data into training and test sets. The purpose of partitioning is to provide one large set of data that can be used to train the regression trees and another smaller one that can be used to evaluate its accuracy.

set.seed(123)
nrow(abalone)

## [1] 4177

i = sample(1:nrow(abalone), 100,replace=FALSE)
abalone.train = abalone[-i,]
abalone.test = abalone[i,]

Now we will specify the gradient boosted model with:

• Poisson distribution. This is not the same as the usual ‘family.’ It does not directly nominate a modelling family. Instead it determines the loss function that is used to evaluate the tree splits.
• var.monotone: indicates which of the predictors should be constrained to monotonic trends (0: not monotonic, 1: positive monotonic, -1: negative monotonic)
• n.trees: the total number of trees to sequentially fit
• interaction.depth: the number of splits within each tree. This roughly equates to the order of interactions. The higher the number, the more complex the interactions can be, however, the faster the potential rate of learning
• bag.fraction: random fraction of data used to fit a single tree
• shrinkage: rate of learning. The smaller the value, the slower the learning rate. Typically, values should be between 0.001 and 0.01.
• train.fraction: the proportion of the data used to train the tree. This is for the purpose of determining the balance between over and underfitting and evaluate the accuracy of the fit, however for the former purpose, this is a cruder approach than either out-of-bag or cross validation. Therefore, it is arguably better to set this fraction to 1 (use 100 percent of the data for training) and use either out-of-bag or cross validation and (as we have done here), withhold a fraction of the full data and perform our own accuracy checking.
• cv.folds: the number of cross validation folds.

The values of n.trees, interaction.depth and shrinkage used below are purely based on what are typically good starting points. Nevertheless, they will be data set dependent. We will start off with those values and then evaluate whether they are appropriate.

abalone.gbm = gbm(RINGS ~ SEX + LENGTH + DIAMETER + HEIGHT +
                      WHOLE_WEIGHT + MEAT_WEIGHT + GUT_WEIGHT +
                      SHELL_WEIGHT,
                  data=abalone.train,
                  distribution='poisson',
                  var.monotone=c(0,1,1,1,1,1,1,1),
                  n.trees=10000,
                  interaction.depth=5,
                  bag.fraction=0.5,
                  shrinkage=0.01,
                  train.fraction=1,
                  cv.folds=3)

We will now determine the optimum number of trees estimated to be required in order to achieve a balance between bias (biased towards the exact observations) and precision (variability in estimates). Ideally, the optimum number of trees should be close to 1000. If it is much less (as in this case), it could imply that the tree learned too quickly. On the other hand, if the optimum number of trees is very close to the total number of fitted trees, then it suggests that the optimum may not actually have occured yet and that more trees should be used (or a faster learning rate).

(best.iter = gbm.perf(abalone.gbm,method='OOB'))

## [1] 264
## attr(,"smoother")
## Call:
## loess(formula = object$oobag.improve ~ x, enp.target = min(max(4, 
##     length(x)/10), 50))
## 
## Number of Observations: 10000 
## Equivalent Number of Parameters: 39.99 
## Residual Standard Error: 8.359e-05

(best.iter = gbm.perf(abalone.gbm,method='cv'))

## [1] 378

Conclusions:

• both out-of-bag and cross validation suggest that the total number of trees greatly exceeds the number necessary to achieve the optimum balance between bias and precision.
• it might be worth decreasing the learning rate.
• we can probably also reduce the total number of trees (so that it runs a little quicker)

abalone.gbm = gbm(RINGS ~ SEX + LENGTH + DIAMETER + HEIGHT +
                    WHOLE_WEIGHT + MEAT_WEIGHT + GUT_WEIGHT +
                    SHELL_WEIGHT,
                  data=abalone.train,
                  distribution='poisson',
                  var.monotone=c(0,1,1,1,1,1,1,1),
                  n.trees=5000,
                  interaction.depth=5,
                  bag.fraction=0.5,
                  shrinkage=0.005,
                  train.fraction=1,
                  cv.folds=3)

(best.iter = gbm.perf(abalone.gbm,method='OOB'))

## [1] 680
## attr(,"smoother")
## Call:
## loess(formula = object$oobag.improve ~ x, enp.target = min(max(4, 
##     length(x)/10), 50))
## 
## Number of Observations: 5000 
## Equivalent Number of Parameters: 39.99 
## Residual Standard Error: 1.247e-05

(best.iter = gbm.perf(abalone.gbm,method='cv'))

## [1] 777

Conclusions:

• both out-of-bag and cross validation suggest that the total number of trees greatly exceeds the number necessary to achieve the optimum balance between bias and precision.
• it might be worth decreasing the learning rate.

Explore relative influence

If a predictor is an important driver of the patterns in the response, then many of the tree splits should feature this predictor. It thus follows that the number of proportion of total splits that features each predictor will be a measure of the relative influence of each of the predictors.

summary(abalone.gbm, n.trees=best.iter)

Conclusions:

• shell weight is overwelmingly the most influential predictor.
• we can also ascribe a kind of signficance to the influence values by determining which of them are greater than \(100/p\) where \(p\) is the number of predictors. The logic here is that if all predictors were equally influential, then they would all have a relative influence of \(100/p\). Hence, any predictors that have a relative influence greater than this number must be explaining more than its share (and more than pure chance) and predictors with relative influence lower are explaining less than chance (and therefore not significantly influential).

Explore partial effects

attr(abalone.gbm$Terms,"term.labels")

## [1] "SEX"          "LENGTH"       "DIAMETER"     "HEIGHT"       "WHOLE_WEIGHT"
## [6] "MEAT_WEIGHT"  "GUT_WEIGHT"   "SHELL_WEIGHT"

plot(abalone.gbm, 8, n.tree=best.iter)

abalone.gbm %>% pdp::partial(pred.var='SHELL_WEIGHT',
                        n.trees=best.iter,
                        recursive=FALSE,
                        inv.link=exp) %>%
    autoplot()

Recursive indicates that a weighted tree traversal method described by Friedman 2001 (which is very fast) should be used (only works for gbm). Otherwise a slower brute force method is used. If want to back transform - need to use brute force.

nms <- attr(abalone.gbm$Terms,"term.labels")
p <- vector('list', length(nms))
names(p) <- nms
for (nm in nms) {
  print(nm)
  p[[nm]] <- abalone.gbm %>% pdp::partial(pred.var=nm,
                                 n.trees=best.iter,
                                 inv.link=exp,
                                 recursive=FALSE,
                                 type='regression') %>%
    autoplot() + ylim(0, 20)
}

## [1] "SEX"
## [1] "LENGTH"
## [1] "DIAMETER"
## [1] "HEIGHT"
## [1] "WHOLE_WEIGHT"
## [1] "MEAT_WEIGHT"
## [1] "GUT_WEIGHT"
## [1] "SHELL_WEIGHT"

do.call('grid.arrange', p)

We might also want to explore interactions…

abalone.gbm %>% pdp::partial(pred.var=c('SHELL_WEIGHT'),
                             n.trees=best.iter, recursive=FALSE, inv.link=exp) %>%
    autoplot()

abalone.gbm %>% pdp::partial(pred.var=c('SHELL_WEIGHT','MEAT_WEIGHT'),
                             n.trees=best.iter, recursive=TRUE) %>%
    autoplot()

g1 = abalone.gbm %>% pdp::partial(pred.var='SHELL_WEIGHT', n.trees=best.iter,
                            recursive=FALSE,inv.link=exp) %>%
    autoplot
g2 = abalone.gbm %>% pdp::partial(pred.var='HEIGHT', n.trees=best.iter,
                            recursive=FALSE,inv.link=exp) %>%
    autoplot


g1 + g2

Explore accuracy

abalone.acc <- abalone.test %>%
  bind_cols(Pred = predict(abalone.gbm,newdata=abalone.test, n.tree=best.iter, type='response'))

with(abalone.acc,  cor(RINGS, Pred))

## [1] 0.6757684

abalone.acc %>%
  ggplot() +
  geom_point(aes(y=Pred,  x=RINGS))

Explore interactions

Computes Friedman’s H-statistic to assess the strength of variable interactions. This measures the relative strength of interactions in models It is on a scale of 0-1, where 1 is very strong interaction In y=β_0+β_1x_1+β_2x_2+β_3x_3.. H= If both main effects are weak, then the H- stat will be unstable.. and could indicate a strong interaction.

One at a time

What were the strong main effects: - SHELL_WEIGHT - HEIGHT - SEX

attr(abalone.gbm$Terms,"term.labels")

## [1] "SEX"          "LENGTH"       "DIAMETER"     "HEIGHT"       "WHOLE_WEIGHT"
## [6] "MEAT_WEIGHT"  "GUT_WEIGHT"   "SHELL_WEIGHT"

interact.gbm(abalone.gbm, abalone,c(1,4), n.tree=best.iter)

## [1] 0.4870872

interact.gbm(abalone.gbm, abalone,c(4,8), n.tree=best.iter)

## [1] 0.06481342

interact.gbm(abalone.gbm, abalone,c(1,4,8), n.tree=best.iter)

## [1] 0.01738704

abalone.gbm %>% pdp::partial(pred.var=c(1),  n.trees=best.iter, recursive=FALSE) %>% autoplot

abalone.gbm %>% pdp::partial(pred.var=c(1, 4),  n.trees=best.iter, recursive=FALSE) %>% autoplot

abalone.gbm %>% pdp::partial(pred.var=c(1, 8),  n.trees=best.iter, recursive=FALSE) %>% autoplot

#plot(abalone.gbm, c(1,4), n.tree=best.iter)
#plot(abalone.gbm, c(5,6), n.tree=best.iter)

abalone.grid = plot(abalone.gbm, c(1,4), n.tree=best.iter, return.grid=TRUE)
head(abalone.grid)

ggplot(abalone.grid, aes(y=HEIGHT, x=SEX)) +
    geom_tile(aes(fill=y)) +
    geom_contour(aes(z=y)) +
    scale_fill_gradientn(colors=heat.colors(10))

All two way interactions

## [1] "i= 1  Name =  SEX"
## [1] "j= 2  Name =  LENGTH"
## [1] "i= 1  Name =  SEX"
## [1] "j= 3  Name =  DIAMETER"
## [1] "i= 1  Name =  SEX"
## [1] "j= 4  Name =  HEIGHT"
## [1] "i= 1  Name =  SEX"
## [1] "j= 5  Name =  WHOLE_WEIGHT"
## [1] "i= 1  Name =  SEX"
## [1] "j= 6  Name =  MEAT_WEIGHT"
## [1] "i= 1  Name =  SEX"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 1  Name =  SEX"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 3  Name =  DIAMETER"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 4  Name =  HEIGHT"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 5  Name =  WHOLE_WEIGHT"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 6  Name =  MEAT_WEIGHT"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 2  Name =  LENGTH"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 3  Name =  DIAMETER"
## [1] "j= 4  Name =  HEIGHT"
## [1] "i= 3  Name =  DIAMETER"
## [1] "j= 5  Name =  WHOLE_WEIGHT"
## [1] "i= 3  Name =  DIAMETER"
## [1] "j= 6  Name =  MEAT_WEIGHT"
## [1] "i= 3  Name =  DIAMETER"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 3  Name =  DIAMETER"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 4  Name =  HEIGHT"
## [1] "j= 5  Name =  WHOLE_WEIGHT"
## [1] "i= 4  Name =  HEIGHT"
## [1] "j= 6  Name =  MEAT_WEIGHT"
## [1] "i= 4  Name =  HEIGHT"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 4  Name =  HEIGHT"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 5  Name =  WHOLE_WEIGHT"
## [1] "j= 6  Name =  MEAT_WEIGHT"
## [1] "i= 5  Name =  WHOLE_WEIGHT"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 5  Name =  WHOLE_WEIGHT"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 6  Name =  MEAT_WEIGHT"
## [1] "j= 7  Name =  GUT_WEIGHT"
## [1] "i= 6  Name =  MEAT_WEIGHT"
## [1] "j= 8  Name =  SHELL_WEIGHT"
## [1] "i= 7  Name =  GUT_WEIGHT"
## [1] "j= 8  Name =  SHELL_WEIGHT"

Tuning

The takes a long time - do over a break

abalone.gbm1 <- gbm.step(data=abalone %>% as.data.frame, gbm.x=1:8, gbm.y=9,
                        tree.complexity=5,
                        learning.rate=0.001,
                        bag.fraction=0.5,
                        n.trees=10000,
                        family='poisson')

## 
##  
##  GBM STEP - version 2.9 
##  
## Performing cross-validation optimisation of a boosted regression tree model 
## for RINGS and using a family of poisson 
## Using 4177 observations and 8 predictors 
## creating 10 initial models of 10000 trees 
## 
##  folds are unstratified 
## total mean deviance =  0.991 
## tolerance is fixed at  0.001 
## ntrees resid. dev. 
## 10000    0.3979 
## now adding trees...

## 
## mean total deviance = 0.991 
## mean residual deviance = 0.337 
##  
## estimated cv deviance = 0.398 ; se = 0.012 
##  
## training data correlation = 0.796 
## cv correlation =  0.746 ; se = 0.01 
##  
## elapsed time -  0.05 minutes 
## 
##  ########### warning ########## 
##  
## maximum tree limit reached - results may not be optimal 
##   - refit with faster learning rate or increase maximum number of trees

summary(abalone.gbm1)

gbm.plot(abalone.gbm1, n.plots=8, write.title = FALSE)

gbm.plot.fits(abalone.gbm1)

find.int <- gbm.interactions(abalone.gbm1)
summary(find.int)

##              Length Class      Mode   
## rank.list     5     data.frame list   
## interactions 64     -none-     numeric
## gbm.call     21     -none-     list

find.int$rank.list

gbm.perspec(abalone.gbm1,6,5)

Random Forest

library(randomForest)
abalone.rf = randomForest(RINGS ~ SEX + LENGTH + DIAMETER + HEIGHT +
                      WHOLE_WEIGHT + MEAT_WEIGHT + GUT_WEIGHT + SHELL_WEIGHT,
                      data=abalone, importance=TRUE,
                      ntree=1000)
abalone.imp = randomForest::importance(abalone.rf)
## Rank by either:
## *MSE (mean decrease in accuracy)
## For each tree, calculate OOB prediction error.
## This also done after permuting predictors.
## Then average diff of prediction errors for each tree
## *NodePurity (mean decrease in node impurity)
## Measure of the total decline of impurity due to each
## predictor averaged over trees
100*abalone.imp/sum(abalone.imp)

##                 %IncMSE IncNodePurity
## SEX          0.12854075      2.955944
## LENGTH       0.05797647      8.461982
## DIAMETER     0.06420550     10.355279
## HEIGHT       0.08923043     13.033870
## WHOLE_WEIGHT 0.06532046     14.727181
## MEAT_WEIGHT  0.13599780     13.665770
## GUT_WEIGHT   0.07368945     12.105419
## SHELL_WEIGHT 0.10624192     23.973352

varImpPlot(abalone.rf)

## use brute force
abalone.rf %>% pdp::partial('SHELL_WEIGHT') %>% autoplot